Why AI Coding Agents Fail at Coordination and How Leases Fix It

AI coding agents capable of writing code and opening pull requests frequently fail not due to reasoning errors but due to coordination breakdowns when multiple agents share the same task environment. Common failure modes include agents crashing mid-task, losing context mid-flight as their context windows fill up, and two agents independently picking up the same task simultaneously. Traditional project trackers like Jira or GitHub Issues were designed for human teams and assume workers will reliably update and clean up their own task states, an assumption that breaks down with unreliable AI processes. A proposed solution, drawn from an open-source tool called rhizome-mcp, replaces permanent task assignment with time-limited leases tied to individual work attempts. Under this model, an agent actively claims a task and must send periodic heartbeats to renew its lease, so if it crashes or goes silent, the lease automatically expires and the task becomes available again without manual intervention.